ENH: add integer-parameter signature in numpy.h

[fbourgey]

Reference issue

What does this implement/fix?

Following #209, I was trying to use Jacobi and Shifted Jacobi polynomials from XSF. It appears those ufunc signatures will be useful for scipy/special/_special_ufuncs.cpp for Jacobi and the following polynomials.

Additional information

AI Generation Disclosure

Codex reviewed.

[dschmitz89]

Ay, didn't expect this. @khushi-411 Was there a hard requirement why we need std::ptrdiff_t instead of a regular int for the Jacobi functions?

[khushi-411]

Hi, @dschmitz89, @fbourgey! The main reason was that the windows tests were failing in CI when I used a test case with an input value of dtype long so I used std::ptrdiff_t.

From here: #209 (comment)

On 64-bit systems std::ptrdiff_t represents 64 bits whereas long is still 32-bit but for 32-bit systems both represents 32-bits.

Reference: #53 (comment)

[dschmitz89]

Hi, @dschmitz89, @fbourgey! The main reason was that the windows tests were failing in CI when I used a test case with an input value of dtype long so I used std::ptrdiff_t.

From here: #209 (comment)

On 64-bit systems std::ptrdiff_t represents 64 bits whereas long is still 32-bit but for 32-bit systems both represents 32-bits.

Reference: #53 (comment)

Thanks for the explanation. To me this looks good then but it would be great if @steppi could take a final look.

[fbourgey]

Hi, @dschmitz89, @fbourgey! The main reason was that the windows tests were failing in CI when I used a test case with an input value of dtype long so I used std::ptrdiff_t.
From here: #209 (comment)

On 64-bit systems std::ptrdiff_t represents 64 bits whereas long is still 32-bit but for 32-bit systems both represents 32-bits.
Reference: #53 (comment)

Thanks for the explanation. To me this looks good then but it would be great if @steppi could take a final look.

@steppi WDYT?

[fbourgey]

Why not simply use int instead of std::ptrdiff_t? The following fbourgey@4814107 commit seems to pass CI.

[khushi-411]

@fbourgey, I thought since int is always 32-bit on both 32-bit and 64-bit systems, so I thought std::ptrdiff_t would be a better choice. What do you think?

[fbourgey]

@steppi can you chime in here?

[steppi]

I think it's fine to do this on a temporary basis. Longer term, we should move to only using fixed width types, using templates in the kernels for integer args to support "generic integer arguments", and to have int32 and int64 versions of the loops whenever there are integer args. I'm OK with merging this though.

[dschmitz89]

I think it's fine to do this on a temporary basis. Longer term, we should move to only using fixed width types, using templates in the kernels for integer args to support "generic integer arguments", and to have int32 and int64 versions of the loops whenever there are integer args. I'm OK with merging this though.

@steppi Would you have time to open a PR that show cases the template approach for the Jacobi polynomials? If that is the preferred route we should use it now already if it is not too complicated.